AGENT SKILL

arbeitsplan-preflight

Read-only inventory of what a swarm run can and cannot do in this repository before any tokens are spent. Use before arbeitsplan-compile or arbeitsplan-run, when the user asks whether this repo is ready for a swarm, or when a run failed and the cause might be the environment rather than the work. Reports git/worktree support, test commands, other plugins' live guards, and any open run lock. Writes nothing.

License 许可
MIT
所属仓库
Anselmoo/werkstoff
最近更新
2026-11-02
Stars
★ 1
Forks
0

01 前言 Front Matter

name
arbeitsplan-preflight
description
Read-only inventory of what a swarm run can and cannot do in this repository before any tokens are spent. Use before arbeitsplan-compile or arbeitsplan-run, when the user asks whether this repo is ready for a swarm, or when a run failed and the cause might be the environment rather than the work. Reports git/worktree support, test commands, other plugins' live guards, and any open run lock. Writes nothing.

03 SKILL.md 正文

Preflight

Know what is on the bench before measuring it. A run that discovers mid-swarm that the repo cannot make a worktree has already spent N dispatches on a result it cannot use.

Steps

  1. Git and worktrees. Is this a git repository? Can it create a worktree here — is the candidate directory ignored? A repo that cannot make worktrees can still use the matrix backend, so report the alternative rather than just failing.

  2. Acceptance checks. What test command does this repo actually use? A run whose criteria carry no runnable check cannot be refereed, and that is better known now.

  3. Other plugins' live guards. Report every guard that would see this run's edits — takt, cupertino, lehre, andon, zeugnis, befund, nacharbeit — and its marker. A fix pass runs under them, and a denial from one is that plugin doing its job.

    Never suggest another plugin's escape hatch. Naming it as an inventory fact is reporting; recommending it is helping the user disable someone else's guard.

  4. Existing state. Is analysis/arbeitsplan/run_scope.json open, and how old? Is .claude/takt.local.md present, and does it carry this plugin's provenance key?

  5. The matrix backend. Is claude on PATH? Is timeout or gtimeout available (cells run untimed without one)? Is this session nested — because if so, the matrix runner will refuse, correctly, and the user must run it from a terminal.

  6. Toolchains (#107). When a design exists, compile_spec.py --design --spec <design> prints one TOOLCHAIN <name> <probe> line per runtime its commands name — script nodes, acceptance, setup, gates, smoke. Run each probe; exit 0 means installed, whatever it prints (go version and java -version share no output convention). A missing toolchain makes the design not ready. For a wave design, also report the helper runtime: a Python

    = 3.11 (python3, python or py -3) that install_waves.py will resolve, and whether this checkout is the primary one (git rev-parse --git-dir equals --git-common-dir) — a wave run launched from a linked worktree halts at its first node.

  7. Say what the run can and cannot measure. A missing checker means the phases that need it are listed as skipped, never silently passed. Name them.

Rules

  • Writes nothing. Not even a state directory.
  • Never infers a missing checker as present. If there is no test command, say the acceptance criteria will have to be structural, and say what that costs.
  • Refuse to go further on an open lock. Hand the user the release command. A stale lock denies every edit, which is the safe direction, and only a person decides to release it.

Output format

arbeitsplan preflight — /Users/x/proj
  git                 yes (worktrees supported, .arbeitsplan/ is gitignored)
  test command        go test ./...        (from the Makefile's `test` target)
  claude on PATH      yes
  timeout             gtimeout (coreutils)
  nested session      YES — the matrix runner will refuse; run it from a terminal
  toolchains          go (go version) ok · shell (sh -c true) ok · julia (julia --version) MISSING
  helper runtime      python3 3.12.3 (>= 3.11) · primary checkout: yes

other guards live in this repository (a run executes under them; a denial from one is
reported, never bypassed):
  cupertino    marker .cupertino
  takt         marker .claude/takt.local.md  (no arbeitsplan provenance key — hand-written)

run lock: none open
state: workflow.json=absent, run_scope.json=absent, matrix results=absent

Can measure: every phase. Cannot measure: nothing skipped.
Design not ready: julia is named by node conform-jl and is not installed here.
Caution: takt.local.md is hand-written. arbeitsplan-compile will REFUSE to overwrite it.
Next: arbeitsplan-compile to turn a problem into a spec.

Resources

  • references/matrix-schema.md — why a nested session cannot run the matrix.
  • references/workflow-spec-schema.md — what compile will need from this repository.
  • references/design-table-schema.md — the toolchains a design names, and how each is probed.

04 同仓库更多 Skills